Qualitrol Shines at the Coil and Winding Tradeshow, Berlin: Strengthening Connections and Showcasing Innovation
Empowering Community Impact - A Recap of Qualitrol's Day of Caring
Pilot Testing of AI Algorithm for Fault Record Analysis
This blog article summarizes a groundbreaking paper recently presented at the PAC World Americas conference, an annual event bringing together experts in Protection, Automation, and Control systems from around the globe. These experts convene to discuss advances in technology, share ideas, and address the continually evolving landscape of their specialized field.
The authors of the paper are esteemed professionals in their respective domains: Finlay Macleod, System Performance Lead Engineer at Scottish Power; Mark Diamond, Product Manager at Qualitrol; Aaron Acton, Senior Data Scientist at Fortive Corporation; and David Cole, Technical Application Specialist at Qualitrol. Each contributor brings a unique perspective and expertise, creating a robust foundation for the research undertaken.
Their collaborative paper, “Pilot Testing of AI Algorithm for Fault Record Analysis,” explores a revolutionary approach to fault record analysis for Scottish Power. The traditional method, a labor-intensive manual process, is given a fresh, tech-forward makeover using AI algorithms. The outcome from the pilot test suggests promising future applications for the AI algorithm in this field, bringing tangible benefits in efficiency and productivity.
Please read on for a summary of the paper’s key findings and insights.
Introduction
A pilot test of an AI algorithm has been conducted to categorize fault types and determine fault causes in fault records for Scottish Power. The traditional process of analyzing fault records involves manually examining the data, which can be time-consuming. Automating the process through AI algorithms would speed up the analysis, leading to a quicker response in dispatching patrol teams and allowing analysts to examine more records.
Scottish Power’s Monitoring System
Scottish Power has a large number of fault recorders that report to a central master station. Many fault records are generated daily, particularly during storm conditions. Analyzing these records is vital for understanding asset health. The pilot test aimed to utilize AI algorithms to categorize fault types, such as trip, through fault, voltage dip, and close onto fault, among others. This would enable analysts to filter and prioritize records for increased productivity.
AI Algorithm and Methodology
The AI algorithm used a Random Forest technique trained on a large dataset of labeled records provided by multiple utilities. Scottish Power set up a separate master station to access 23 fault recorders in 13 substations for the pilot test. The devices were chosen for their likelihood of producing the maximum number of fault records in a short time. The results have been more successful than previous rules-based analysis, and the trial will be extended to more substations in the future.
Scottish Power’s Transmission Network
Scottish Power operates in the UK and is responsible for transmitting electricity generated from power stations, wind farms, and other utilities through a large transmission network. The company has been actively installing and operating a monitoring program for over 25 years, with a dedicated team assigned to oversee the program and analyze the data.
Benefits of Automating Fault Record Analysis
The benefits of automating the analysis of fault records include providing information on protection performance, identifying circuit breaker issues, offering advance warning of voltage transformer problems, and assessing voltage levels and power quality at supply points.
Development of Machine Learning Algorithm
Developing a machine learning algorithm to classify fault records by category and root cause was the primary objective of the pilot test. The algorithm was trained on a large set of historical records collected from different transmission utilities and stored in Comtrade format. The initial model achieved a 90% accuracy across 10 fault categories and reached 97% after several refinement iterations.
Conclusion
The success of the machine learning algorithm in detecting record categories compared to previous rules-based methods means progress can now be made on a record analyses package that better serves the needs of data analysts. The potential for identifying root cause is another important feature to determine the next steps in remedial actions that will contribute to reducing downtime and assist in the preparation of statistics for submission to the regulator.
